5b03e25b

test: e2e — (a) asserts the second marker ABSENT before ssh dies

a73x   2026-08-11 18:46

test: e2e — (a) asserts the second marker ABSENT before ssh dies

Correction to bde1af5, from the spec review. That commit added the
post-hoc half of the ordering proof and called it done; the half that
actually closes the hole was missing.

The hole had a real failure mode behind it. Both markers are typed by a
script running on its own clock, in parallel with the attach. If an attach
ever took longer than the 6s between them, both would be typed before the
session existed, both would be delivered out of the carry buffer the
moment it did, and cold-two would already be in the capture when the
client killed ssh. Every check in the scenario would then pass — pid
observed dead, marker observed present, pid still dead — while proving
nothing whatsoever about bytes moving after the pipe's owner died.

So the marker is now asserted ABSENT immediately before the death is
observed. That fixes the ORDER; the post-hoc kill -0 fixes the DURATION
and rules out pid reuse. Each comment names which property it carries,
because neither implies the other and a reader who thinks it does will
delete the wrong one.

A failure of the new check is this scenario's timing premise breaking —
an attach slower than 6s against ~300ms observed — not a product defect,
and the message says so rather than blaming the code.

Checked in situ rather than assumed: pointed at "cold-one", which IS
present at that moment, the check fires with its own message and the run
goes red. Restored to "cold-two", the suite is green at 20 scenarios / 33
convergence points.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>